You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Closing as not reproducible. Recording what was checked so this can be reopened with one file attached.
What the screenshot shows. The dialog is the ChatGPT desktop app's own crash panel (「ChatGPT 意外停止」 / 更新 ChatGPT / 打开 Config.toml / 重启), not a codex CLI traceback. So the report is "the Codex host process died while a Strict review was running", which is a crash on the reporter's machine, in a GUI application, with no stack trace, no log excerpt, and no codex/app version recorded.
Why no fix is being guessed at. Every mitigation one would reach for was already in place before this was filed (2026-08-15):
Strict runs in an isolated subprocess, not in the server process — mcp_server.py, start_new_session=True on POSIX, CREATE_NEW_PROCESS_GROUP | CREATE_NO_WINDOW on Windows, stdin=subprocess.DEVNULL (all since 2026-07-22).
The MCP server ignores inherited console SIGINT on Windows, so a host Ctrl+C event cannot tear the run down.
Strict has been asynchronous since 2026-07-30: it returns a run_id immediately and is polled, so the host is never blocked waiting on a long review.
Run state is durable (run_status.json) with ownership-aware reconciliation, so a dead host does not lose or duplicate a run.
The crash happened anyway, which means the cause is not any of these, and nothing in this report distinguishes between the remaining candidates. Shipping a speculative "fix" against an unreproduced GUI crash would close the issue without changing anything.
It also cannot be reproduced here. The ChatGPT desktop app is not installed on the maintenance host, and the Strict backend has since moved to cursor; the codex Strict backend is still unlanded (#82).
To reopen with a diagnosis attached, please include:
~/.codex/log/ from around the crash (or the app's crash report),
the ChatGPT app version and codex --version,
the OS, and
whether it reproduces with a second Strict run on the same PR.
That turns this into something actionable. Without it there is nothing to act on, and an open issue that nobody can act on is worse than a closed one that can be reopened in a minute.
Activity
Merge pull request #103 from JiusiServe/fix/thin-mcp-tool-annotations
tzhouam commentedon Sep 11, 2026
Closing as not reproducible. Recording what was checked so this can be reopened with one file attached.
What the screenshot shows. The dialog is the ChatGPT desktop app's own crash panel (「ChatGPT 意外停止」 / 更新 ChatGPT / 打开 Config.toml / 重启), not a
codexCLI traceback. So the report is "the Codex host process died while a Strict review was running", which is a crash on the reporter's machine, in a GUI application, with no stack trace, no log excerpt, and nocodex/app version recorded.Why no fix is being guessed at. Every mitigation one would reach for was already in place before this was filed (2026-08-15):
mcp_server.py,start_new_session=Trueon POSIX,CREATE_NEW_PROCESS_GROUP | CREATE_NO_WINDOWon Windows,stdin=subprocess.DEVNULL(all since 2026-07-22).run_idimmediately and is polled, so the host is never blocked waiting on a long review.run_status.json) with ownership-aware reconciliation, so a dead host does not lose or duplicate a run.The crash happened anyway, which means the cause is not any of these, and nothing in this report distinguishes between the remaining candidates. Shipping a speculative "fix" against an unreproduced GUI crash would close the issue without changing anything.
It also cannot be reproduced here. The ChatGPT desktop app is not installed on the maintenance host, and the Strict backend has since moved to
cursor; thecodexStrict backend is still unlanded (#82).To reopen with a diagnosis attached, please include:
~/.codex/log/from around the crash (or the app's crash report),codex --version,That turns this into something actionable. Without it there is nothing to act on, and an open issue that nobody can act on is worse than a closed one that can be reopened in a minute.
🤖 Generated with Claude Code
https://claude.ai/code/session_01KT4EvUTzzLeq243DG6T1mC